perf(vector/search): stop materializing vector_ids() on count and warmup paths (#672)

[mosuka]

Summary

Since #633 the audit-flagged "clone" in vector_ids() is really an O(N) per-call String rehydration at the trait boundary (readers store interned (u64, u16) + a field dictionary). Rather than breaking the public vector_ids() signature, this PR stops every per-query and startup call site from using it:

• Count paths (hnsw/flat/ivf searchers, no-field branch): .len() of a full materialization → vector_count() (equal by construction: one entry per (doc, field) record).
• SegmentFanoutSearcher::count: per-segment full materialization → per-field iteration over doc_ids_for_field (an O(1) Arc clone on all three reader types, perf(vector): cache vector_ids per field at reader load #405), with dictionary-backed field_names() supplying the field list for unfiltered counts. Newest-wins masking and deletion filtering unchanged.
• HNSW warmup (OnDemand page-fault pass): new pub(crate) HnswIndexReader::interned_vector_ids() borrows dictionary names — the code already held the concrete reader, so the 10M-String startup spike disappears with no trait change.
• Left as-is, documented (per the plan comment): the unfiltered brute-force fallbacks and facade construction, where the materialization immediately precedes an O(N) scan that reads every vector.

Measured

10 segments × 2k docs, 100 count calls, release, best of 3: count(field) 380.7 → 352.6ms, count(None) 378.5 → 350.4ms (-7.4% both). Honest read: at this in-memory scale the containment masking dominates and the allocation was ~7% — the real win is that the removed allocation grows linearly with corpus size (~32 B/record/segment; the audit priced 320 MB per call at 10M records), so a scale-proportional allocation spike is gone rather than a large steady-state percentage.

Tests

• New count_masks_duplicates_and_deletions regression test on the bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927 mixed-range fixture: distinct live (doc, field) keys, cross-segment duplicates counted once, deletions excluded, field filter honored (missing field → 0) — count semantics pinned unchanged across the rewrite.
• Existing count/warmup/fan-out tests green unchanged.

Verification

• cargo fmt --check and cargo clippy --all-targets -- -D warnings clean on stable and 1.97.0
• cargo test -p laurus --lib: 1258 passed; --tests: all 64 binaries ok
• wasm32 cargo check -p laurus-wasm --target wasm32-unknown-unknown: clean

Closes #672. Builds on #633 (interning) / #405 (doc_ids_for_field) / #926 (shared fan-out).